MSTAF-CDN: Multi-Scale Temporal Attention Fusion Network for Remote Sensing Change Detection

Deep learning has profoundly advanced remote sensing change detection (RSCD). However, effectively modeling the temporal interdependencies between bi-temporal images remains a challenge, often resulting in inadequate feature fusion, blurred change boundaries, and missed detections of subtle changes. To tackle these issues, we propose Multi-Scale Temporal Attention Fusion Change Detection Network(MSTAF-CDN), a novel network that pioneers a multi-scale temporal attention fusion strategy. This network with the Two-stage Collaborative Attention Assembly (TCAA) module, which explicitly captures cross-temporal contextual relationships to guide the dynamic fusion of multi-scale features, and the Progressive Contour Sensing Augmentation (MFSA) module, which employs multi-branch architecture to progressively sharpen boundary details and enhance sensitivity to small targets. Improve separately from the intrinsic challenges at the data level and the deep requirements of temporal modeling. On three public benchmarks, MSTAF-CDN achieves competitive results, with F1 scores of 91.54% on LEVIR-CD, 83.70% on SYSU-CD, and 76.27% on UAV-CD, outperforming several recent state-of-the-art methods, demonstrating superior performance in both quantitative metrics and visual quality. Our code is available at https://github.com/yikuizhai/MSTAF-CDN.